MORE: A Framework for Stable White Blood Cell Morphological Classification and Report Generation

In hematological diagnostics, the morphological assessment of blood cells serves as a fundamental cornerstone, where clinicians manually evaluate attributes like cell shape, size, and color to identify cell types and detect disease-indicating abnormalities, particularly in conditions such as anemia and leukemia. Automatic analysis accelerates diagnostic throughput, minimizes human error, and supports earlier detection of diseases. To address these challenges, we present MORE, a novel end-to-end framework composed of two main components: (1) an image captioning system built upon an ensemble of ten lightweight backbone classifiers that leverages both soft and hard voting strategies to handle class imbalance and batch effects, and (2) dual report generation modules that transform the captioning output into clinically meaningful narratives. While previous automated WBC analysis focused solely on type classification without considering morphological attributes or generating human-readable reports, MORE bridges this gap by delivering both accurate classification and interpretable reporting. Testing our framework on the WBCAtt and LeukemiaAttri datasets, the image captioning component achieves up to 6.8% macro-F1 gain over the state-of-the-art on 12 WBCAtt’s WBC attributes, while maintaining competitive performance on LeukemiaAttri. The report generation modules consist of a deterministic engine that ensures consistent clinical summaries and a customizable Llama-3.1-8B narrative generator that provides context-aware insights. By combining robust image analysis with automated report generation, MORE represents the first comprehensive solution for translating WBC images into clinically meaningful documentation, advancing both quantitative analysis and qualitative reporting in hematological diagnostics. The code associated with this manuscript is available at: https://github.com/unica-visual-intelligence-lab/MORE-WBC.